Dynamic windowed codebook search algorithm in vector quantization

Abstract

We present an efficient codebook search algorithm in a vec- tor quantization (VQ)-based system that searches only through a window of consecutive codewords on the preordered codebook, rather than all codewords in the codebook, and achieves full-search-equivalent perfor- mance. The ordering technique used to rearrange the positions of code- words in the codebook utilizes the scattering orientation of codewords in the k-dimensional space or the compactness property of signal energy on the transform domain. The procedure for determining the search win- dow on the ordered codebook relies on the geometric relations among input vector and codewords to eliminate those codewords that are cer- tain not to be the closest codeword to the input vector. Finally, the clos- est codeword to an input vector is selected from the determined search window by an improved partial distortion search algorithm. In compari- son with other existing fast algorithms, the proposed algorithm requires the least number of multiplications and the least total number of distor- tion measurements. © 1996 Society of Photo-Optical Instrumentation Engineers. Subject terms: vector quantization; codebook search; nearest neighbor; data compression; principal components analysis.